If you want to use these in Torch7, my excellent colleague @soumith has released bindings at https://github.com/soumith/cudnn.torch. There are some benchmarks for the convolutional layers of a classic AlexNet model at https://github.com/soumith/convnet-benchmarks.
In general, the question which is the fastest public implementation of spatial convolution is so heavily dependent on the kernel parameters (and even then, there are often knobs you can twiddle for a given implementation that can lead to substantial speedups), how much GPU memory you want to spend (GEMM and FFT based methods), so the best approach is probably just to benchmark a bunch of implementations for your network and just choose the best.